[A-AS05] Weather, Climate, and Environmental Science Studies using High-Performance Computing

[AAS05-09] Tropical cyclone seasonal hindcast by a Neural-Physics hybrid AGCM

*Masuo Nakano^1,2, Yohei Yamada¹, Takeshi Enomoto^3,1, Akira Yamazaki¹ (1.Japan Agency for Marine-Earth Science and Technology, 2.Yokohama National University Typhoon Science and Technology Research Center (TRC), 3.Kyoto University Disaster Prevention Research Institute)

Keywords:seasonal forecast, Neural-Physics hybrid model, Tropical cyclone

Accurate seasonal prediction of tropical cyclones (TCs) is essential for planning to mitigate their impact. The use of high-resolution coupled models is one of the possible ways to improve TC seasonal prediction accuracy. However, such models are computationally expensive. From a predictability perspective, understanding the sources of forecast errors can lead to further improvements in prediction skill. However, due to the presence of model biases, it is challenging to diagnose whether prediction failures stem from the model biases or from the chaotic nature of the atmosphere. The emergence of AI-based global climate models trained on historical atmospheric reanalysis data may help with such diagnoses because they show lower bias than physics-based GCMs. Here, we performed a 10-member ensemble hindcast experiment for 14 boreal TC seasons (JJASO of 2010-2023) using NeuralGCM. The AI-based model well reproduced the 14-season mean TC genesis, track density, and the seasonal march of TC frequency. Furthermore, the AI-based model produced skillful seasonal TC forecasts even for the western North Pacific, where it has long been believed that skillful TC seasonal forecasting is difficult. This result suggests further improvement of traditional physics-based GCMs will lead to more skillful seasonal forecasting of TCs.